1. Distributed Computing Group On the Impact of Malicious Players in Distributed Systems Stefan Schmid DYNAMO’07: 1st Workshop on Dynamic Networks (Salerno, Italy, May 2007) Talk is based on PODC’06 paper, joint work with Dr. Thomas Moscibroda, MS Research Prof. Dr. R. Wattenhofer, ETH Zurich

2. Stefan Schmid @ Dynamo, 20072 Distributed Systems… Wireless Sensor Networks Internet Wireless Mesh Networks P2P Networks (Electronic) Markets, Society,…?

3. Stefan Schmid @ Dynamo, 20073 Modeling Participants of Distributed Systems One possibility to model a distributed system: all participants are benevolent! Network

4. Stefan Schmid @ Dynamo, 20074 Selfishness in Networks Alternative: Model all participants as selfish  e.g. impact on congestion, or on p2p topologies, etc. Network Classic game theory: What is the impact of selfishness on network performance…? (=> Notion of price of anarchy, etc.)

5. Stefan Schmid @ Dynamo, 20075 When Selfish meets Evil… But selfishness is not the only challenge in distributed systems!  Malicious attacks on systems consisting of selfish agents Hackers, Polluters, Viruses, DOS attacks What is the impact of malicious players on selfish systems…? Network

6. Stefan Schmid @ Dynamo, 20076 Goal of a selfish player: minimize her own cost Social Cost is the sum of costs of selfish players Some Definitions from Game Theory Social Optimum (OPT) –Minimal social cost of a given problem instance –“solution formed by collaborating players”! Nash equilibrium –“Result” of selfish behavior –State in which no selfish player can reduce its costs by changing her strategy, given the strategies of the other players Measure impact of selfishness: Price of Anarchy –Captures the impact of selfishness by comparison with optimal solution –Formally: social costs of worst Nash equilibrium divided by optimal social cost Large PoA -> Selfish player are harmful!

7. Stefan Schmid @ Dynamo, 20077 Of course, whether a selfish player is happy with its situation depends on what she knows about the malicious players! Do they know that there are malicious players? If yes, it will take this into account for computing its expected utility! Moreover, a player can react differently to knowledge (e.g. risk averse). “Byzantine* Game Theory” Game framework for malicious players Consider a system (network) with n players Among these players, s are selfish System contains b=n-s malicious players Malicious players want to maximize social cost! Define Byzantine Nash Equilibrium: A situation in which no selfish player can improve its perceived costs by changing its strategy! Social Cost: Sum of costs of selfish players: * „malicious“ is better… but we stick to paper notation in this talk.

8. Stefan Schmid @ Dynamo, 20078 Actual Costs vs. Perceived Costs Depending on selfish players‘ knowledge, actual costs (-> social costs) and perceived costs (-> Nash eq.) may differ! Actual Costs:  The cost of selfish player i in strategy profile a Perceived Costs:  The cost that player i expects to have in strategy profile a, given preferences and his knowledge about malicious players! Nothing…, Number of malicious players… Distribution of malicious players… Strategy of malicious players… Risk-averse… Risk-seeking… Neutral… Many models conceivable Players do not know ! Byz. Nash Equilibrium

9. Stefan Schmid @ Dynamo, 20079 Game theory with selfish players only studies the Price of Anarchy: We define Price of Byzantine Anarchy: Finally, we define the Price of Malice! How to Measure the Impact of Malicious Players? The Price of Malice captures the degradation of a system consisting of selfish agents due to malicious participants! Social Optimum Worst NE Worst NE with b Byz. Price of Anarchy Price of Malice Price of Byzantine Anarchy

10. Stefan Schmid @ Dynamo, 200710 Are malicious players different from selfish players...? Also egoists?! Theoretically, malicious players are also selfish....... just with a different utility function!  Difference: Malicious players‘ utility function depends inversely on the total social welfare! („irrational“: utility depends on more than one player‘s utility)  When studying a specific game/scenario, it makes sense to distinguish between selfish and malicious players. Remark on “Byzantine Game Theory” Everyone is selfish!

11. Stefan Schmid @ Dynamo, 200711 Sample Analysis: Virus Inoculation Game Given n nodes placed in a grid network Each peer or node can choose whether to install anti-virus software Nodes who install the software are secure (costs 1) Virus spreads from one randomly selected node in the network All nodes in the same insecure connected component are infected (being infected costs L, L>1)  Every node selfishly wants to minimize its expected cost! Related Work: The VIG was first studied by Aspnes et al. [SODA’05] General Graphs No malicious players

12. Stefan Schmid @ Dynamo, 200712 What is the impact of selfishness in the virus inoculation game? What is the Price of Anarchy? Intuition: Expected infection cost of nodes in an insecure component A: quadratic in |A| |A|/n * |A| * L = |A| 2 L/n Total infection cost: Total inoculation cost: Virus Inoculation Game: Selfish Players Only A k i : insecure nodes in the ith component  number of secure (inoculated) nodes Optimal Social Cost Price of Anarchy: Simple … in NE, size <n/L+1 otherwise inoculate!

13. Stefan Schmid @ Dynamo, 200713 Adding Malicious Players… What is the impact of malicious agents in this selfish system? Let us add b malicious players to the grid! Every malicious player tries to maximize social cost!  Every malicious player pretends to inoculate, but does not! (worst-case: malicious player cannot be trusted and may say s.th. but do s.th. else…) What is the Price of Malice…?  Depends on what nodes know and how they perceive threat! Distinguish between: Oblivious model Non-oblivious model Risk-averse

14. Stefan Schmid @ Dynamo, 200714 Nodes do not know about the existence of malicious agents (oblivious model)! They assume everyone is selfish and rational How much can the social cost deteriorate…? Simple upper bound: At most every selfish node can inoculate itself  Recall: total infection cost is given by (see earlier: component i is hit with probability k i /n, and we count only costs of the l i selfish nodes therein) Price of Malice – Oblivious case Size of attack component i (including Byz.) #selfish nodes in component i

15. Stefan Schmid @ Dynamo, 200715 Total infection cost is given by: It can be shown: for all components without any malicious node  (similar to analysis of PoA!) On the other hand: a component i with b i >0 malicious nodes: In any non-Byz NE, the size of an attack component is at most n/L, so Price of Malice – Oblivious case it can be shown

16. Stefan Schmid @ Dynamo, 200716 Adding inoculation and infection costs gives an upper bound on social costs: Hence, the Price of Byzantine Anarchy is at most The Price of Malice is at most Price of Malice – Oblivious case for b<L/2 (for other case see paper) Because PoA is if L<n

17. Stefan Schmid @ Dynamo, 200717 In fact, these bounds are tight! I.e., there is instance with such high costs.  bad example: components with large surface (many inoculated nodes for given component size => bad NE! All malicious players together, => and one large attack component, large BNE)  this scenario where every second column is is fully inoculated is a Byz Nash Eq. in the oblivious case, so:  What about infection costs? With prob. ((b+1)n/L+b)/n, infection starts at an insecure or a malicious node of an attack component of size (b+1)n/L  With prob. (n/2-(b+1)n/L)/n, a component of size n/L is hit Oblivious Case Lower Bound: Example Achieving It… 2b n/L Combining all these costs yields

18. Stefan Schmid @ Dynamo, 200718 So, if nodes do not know about the existence of malicious agents! They assume everyone is selfish and rational Price of Byzantine Anarchy is: Price of Malice is: Price of Malice – Oblivious case This was Price of Anarchy… Price of Malice grows more than linearly in b Price of Malice is always ¸ 1  malicious players cannot improve social welfare! This is clear, is it…?!

19. Stefan Schmid @ Dynamo, 200719 Price of Malice – Non-oblivious Case Selfish nodes know the number of malicious agents b (non-oblivious) Assumption: they are risk-averse The situation can be totally different… …and more complicated! For intuition: consider the following scenario…: more nodes inoculated! Each player wants to minimize its maximum possible cost (assuming worst case distribution) n/L This constitutes a Byzantine Nash equilibrium! Any b nodes can be removed while attack component size is at most n/L! (n/L = size where selfish node is indifferent between inoculating or not in absence of malicious players)

20. Stefan Schmid @ Dynamo, 200720 Price of Malice – Lower Bound for Non-oblivious Case What is the social cost of this Byzantine Nash equilibrium…? (all b malicious nodes in one row, every second column fully inoculated, attack size =< n/L) n/L Total inoculation cost: Infection cost of selfish nodes in infected row… n/L-b selfish nodes (b > n/L -> all s nodes inoculate) It can be shown that expected infection cost for this row is: Infection cost of selfish nodes in other rows… number of insure nodes in other rows Total Cost: L

21. Stefan Schmid @ Dynamo, 200721 Price of Malice – Non-oblivious Case: Lower Bound Results Nodes know the number of malicious agents b Assumption: Non-oblivious, risk-averse Price of Byzantine Anarchy is: Price of Malice is: Price of Malice grows at least linearly in b Price of Malice may become less than 1…!!!  Existence of malicious players can improve social welfare! (malicious players cannot do better as we do not trust them in our model, i.e., not to inoculate still is the best thing for them to do!)

22. Stefan Schmid @ Dynamo, 200722 In the non-oblivious case, the presence (or at least believe) of malicious players may improve social welfare! Selfish players are more willing to cooperate in the view of danger! Improved cooperation outweighs effect of malicious attack! In certain selfish systems: Everybody is better off in case there are malicious players! Define the Fear-Factor  The Windfall of Malice: the “Fear Factor”  describes the achievable performance gain when introducing b Byzantine players to the system! In virus game:

23. Stefan Schmid @ Dynamo, 200723 Price of Malice – Interpretations & Implications What is the implication in practical networking…? If Price of Anarchy is high  System designer must cope with selfishness (incentives, taxes) If Price of Malice is high  System must be protected against malicious behavior! (e.g., login, etc.) Price of Anarchy orthogonal Price of Malice <1 >1 1 smalllarge use incentives add malicious agents Protect against malicious agents Use incentives or malicious agents

24. Stefan Schmid @ Dynamo, 200724 Reasoning about the Fear Factor What is the implication in practical networking…? Fear-Factor can improve network performance of selfish systems! (if Price of Malice < 1) Are there other selfish systems with  >1 ? If yes… make use of malicious participants!!! Possible applications in P2P systems, multi-cast streaming, …  Increase cooperation by threatening malicious behavior! In our analysis: we theoretically upper bounded fear factor in virus game!  That is, fear-factor is fundamentally bounded by a constant (independent of b or n)

25. Stefan Schmid @ Dynamo, 200725 Future Work and Open Questions Plenty of open questions and future work! Virus Inoculation Game  The Price of Malice in more realistic network graphs  High-dimensional grids, small-world graphs, general graphs,…  How about other perceived-cost models…? (other than risk-averse)  How about probabilistic models…? The Price of Malice in other scenarios and games  Routing, caching, etc…  Fear-Factor in other systems…?  Can we use Fear-Factor to improve networking…? THANK YOU ! Recent study of congestion games: „Congestion Games with Malicious Players“ by M. Babaioff, R. Kleinberg, C. Papadimitriou (EC’07)